YouTube Transcript Scraper — works on videos with no captions

Most transcript scrapers read the caption track that YouTube already published. When a video has no captions, they return an error.

This one transcribes the audio instead.

How often does that happen? Measured across 172 videos on 2026-07-29 — and it depends entirely on what you are processing:

content	share with no caption track
English	0–10 %
Polish, Turkish, Vietnamese, Thai	~10 %
Romanian	30 %
Czech	50–75 %
Hungarian	80 %
Live concerts and DJ sets	50 %
TikTok, Twitch	100 % — these platforms publish no caption tracks at all

(Vimeo is not in that table: it was not part of the sample, and some Vimeo videos do carry uploader-provided subtitles — those are used when they can be fetched.)

If your inputs are mainstream English videos, a plain caption reader will serve you fine and you do not need this. If they are in a smaller language, are live recordings, or come from anywhere other than YouTube, a caption reader returns nothing for a large share of them.

What you get

{
  "title": "…",
  "durationSeconds": 4022,
  "source": "manual_captions | auto_captions | whisper",
  "isAiGenerated": false,
  "language": "cs",
  "text": "the full transcript…",
  "wordCount": 8213,
  "segments": [{ "start": 0.2, "end": 3.1, "text": "…" }],
  "vtt": "WEBVTT…",
  "lowConfidence": false,
  "warning": null
}

Three things it does that a caption reader cannot

1. It uses the free path when there is one. If the video already has captions, they are read — instantly, and you are not charged for speech recognition you did not need. Whisper runs only when there is nothing to read. Set forceTranscription if you want the audio transcribed anyway.

2. It refuses to hand you a translation and call it a transcript. YouTube machine-translates its own machine transcript into ~150 languages and publishes them all as "captions". Ask for Czech captions on an English video and most tools will happily give you a translation of an ASR guess. This one detects those and transcribes the real audio instead.

3. It tells you when it could not hear anything. Speech-recognition models trained on subtitle files invent subtitle furniture on silence and music — translator credits, "thanks for watching", channel names. Those are stripped, and if little real speech remains the result is flagged lowConfidence with a reason. An honest "we could not hear speech in this" is more useful than a fluent invention, and you can filter on it.

Also works on

TikTok · Twitch VODs · Vimeo · SoundCloud · direct links to audio or video files — about 1 800 sites in total.

TikTok and Twitch publish no caption tracks at all, so everything from them is transcribed. Vimeo sometimes does carry uploader-provided subtitles, and those are used when they can be fetched. A direct link to an .mp3 or .mp4 carries no duration metadata; the file is measured after download rather than refused.

Playlists, channels and collection URLs are refused rather than expanded — one of those can be thousands of videos, and you would be billed for all of them. Send one video per item.

Input

field	what it does
videoUrls	one URL per line. Bare YouTube IDs work. Playlists are refused — they can be thousands of videos and you would be billed for all of them.
language	two-letter code, e.g. en, cs, hu. Leave empty for whatever the video is in.
forceTranscription	transcribe the audio even when captions exist
includeSubtitles	return a ready-to-use WebVTT file
includeCutList	return the spans of dead air, for editors cutting long footage

Notes

Long recordings are cut into overlapping windows, transcribed, and reassembled onto the original clock, so timings stay correct on a two-hour stream. If any window fails, the run reports an error rather than returning a transcript with a silent gap in the middle — a partial transcript reads perfectly and there is no way for you to tell.